Predição genômica ampla de híbridos simples de milho considerando modelo não aditivo

Abstract

It was evaluated by simulation the behavior of genomic prediction models against the inclusion of non-additive effects of dominance and/or epistasis at the prediction of single cross hybrids of maize for the ear weight character. The objectives were to assess G-BLUP models of hybrid prediction with inclusion of non-additive effects and compare with the Bayesian Ridge Regression (BRR), Bayes A, Bayes LASSO, Bayes C, Bayes B and Bayesian Reproducing Kernel Hilbert Spaces Regression (RKHS) models, implemented in the BGLR package of R software, with inclusion or absence of non-additive effect in three scenarios of heritability. It was used data from the maize germplasm bank of the United States Department of Agriculture to find parametric values of variance components, from phenotypic data available, and with it simulated the achievement of 400 single cross hybrids between two groups of lines taken from the whole data available. The models were compared by sum of squares of prediction errors (PRESS), accuracy of prediction, estimation of variance components, correlation between simulated and estimated values and finally heritability of several kinds of genetics effects present in every situation. Altogether, all the models responded positively to the inclusion of non-additive effects, especially dominance. For some comparative criteria, the G-BLUP models performed notably higher to the others, but the opposite was not found for any of the other models. Slight differences among the performance of the models occurred according to the heritability range considered. The worst estimates of genetic effects were always related to the epistatic effects. So, it was concluded that the resulting estimates with the G-BLUP models, mainly, and the other models tested became more accurate when taking into account non-additive effects for a quantitative character as ear weight in maize.